Studying the impact of synchronization frequency on scheduling tasks with dependencies in heterogeneous systems

  • Authors:

  • Theodore Andronikos;Florina M. Ciorba;Ioannis Riakiotakis;George Papakonstantinou;Anthony T. Chronopoulos

  • Affiliations:

  • Department of Informatics, Ionian University, Corfu, Greece;Computing Systems Laboratory, Department of Electrical & Computer Engineering, National Technical University of Athens, Greece and Center for Advanced Vehicular Systems, Mississippi State Universi ...;Computing Systems Laboratory, Department of Electrical & Computer Engineering, National Technical University of Athens, Greece;Computing Systems Laboratory, Department of Electrical & Computer Engineering, National Technical University of Athens, Greece;Department of Computer Science, University of Texas at San Antonio, 6900 N. Loop 1604 West, San Antonio, TX 78249, USA

  • Venue:
  • Performance Evaluation
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

In this work, we develop and evaluate a theoretical model, which we then use to study the impact of the synchronization frequency on the performance of dynamic self-scheduling algorithms. These algorithms are used to parallelize loops with data dependencies on heterogeneous systems. The proposed model uses a formula to estimate the parallel time as a function of the synchronization frequency. Inter-node communication has been proven to be the dominant factor for the performance degradation of applications containing loops with data dependencies. The synchronization mechanism therefore requires careful fine-tuning in order to give the best possible performance. The proposed model determines the optimal synchronization frequency that results in the minimum parallel time. We use this model to study the impact of the synchronization frequency on the parallel execution of a computational kernel from image processing. For this kernel, the synchronization frequency giving the minimum parallel time predicted by our theoretical model was very close to the synchronization frequency giving the least parallel time in practice. We validate our model by extensive comparisons of the theoretically predicted parallel time and synchronization frequency against those obtained from practical experiments. The comparisons show that the proposed model is highly accurate, its predictions for the optimal synchronization frequency being within 0.0250% of the experimentally optimal synchronization frequency in the best case, and within 0.1750% of the experimentally optimal synchronization frequency in the worst case. Finally, the comparisons show that the proposed model improves on a previously existing model in heterogeneous systems, whereas it gives similar results in homogeneous systems.